An improved framework for anchor pixel selection in the Surface Energy Balance Model (SEBAL) for consistent estimation of evapotranspiration

Accurate evapotranspiration (ET) estimation in an energy balance model, such as SEBAL, depends on the selection of hot and cold anchor pixels. This selection is highly subjective and varies among modellers, leading to substantial uncertainty and frequent inconsistencies in ET outputs. Manual identification is time-consuming and often problematic, with incorrect anchor-pixel choices producing unrealistic energy partitioning and even negative ET estimates. The sensitivity analysis in this study revealed that hot anchor pixels with identical land surface temperature (LST) but at different spatial locations can yield significantly different, and sometimes negative, ET estimates, underscoring the need for an optimal, physically representative anchor-pixel pair. Therefore, this study introduces an automated optimization framework using a genetic algorithm that selects an anchor pixel pair from the candidate anchor pixel pool, defined based on the NDVI-LST feature space, to maximize the evaporative fraction over agricultural areas while maintaining energy balance closure. The approach eliminates subjectivity in anchor pixel selection and ensures consistent energy-balance computation. The model performance was evaluated using 10-day cumulative ET estimates and the water balance model, ORYZA2000, which was parameterized with detailed field-level crop, water, soil and meteorological information. Optimized SEBAL estimates showed strong agreement with independent ET estimates from field-parameterized water balance model, with R2 values between 0.89 and 0.94 and RMSE values of 5.21–8.56 mm. In contrast, ET estimates using manual anchor pixel selection exhibited lower performance with R2 values between 0.77 and 0.83 and RMSE values between 7.69 and 10.52 mm. Moreover, the negative ET values obtained during the manual approach were automatically excluded while executing optimized SEBAL. The proposed method is automated and ensures consistent ET estimates, thereby greatly improving the operational reliability and applicability of the SEBAL model across a large spatial domain.